TriNN: A Concise, Lightweight, and Fast Global Triangulation GNN for Point Cloud

In the field of practical applications for point cloud neural networks, besides precision, high real-time performance and low resource utilization often hold significant importance. However, traditional methods such as RNN and k-NN graph construction, often employed in point clouds GNNs, tend to suffer from low real-time performance and high resource consumption. To tackle these challenges, this work introduces a concise, lightweight, and fast global triangulation GNN (TriNN). To replace RNN and k-NN, the Range Plane and Range Belt are proposed for constructing a Delaunay triangulation-based graph on point clouds. Importantly, both the range plane and range belt can be triangulated without relying on point-wise normals. The resulting graph not only encapsulates the raw point cloud in its most concise representation but also preserves all adjacency relationships. Finally, we evaluate the performance of the proposed architecture with respect to overfitting, resource consumption, time cost, and accuracy trade-offs. Experimental results substantiate that TriNN is adept at constructing deeper networks, demands fewer computational resources, and achieves faster computation.